When the algorithm picks the audience, what do we still buy?
Automated campaigns can expand beyond audience suggestions, but audience controls still matter. Meta distinguishes suggestions from controls, and Google Performance Max treats audience signals as guidance. Set required restrictions correctly, then invest in creative, accurate conversion signals and tests that can measure
Automated campaigns can expand beyond audience suggestions, but required audience restrictions still matter. Meta distinguishes Audience controls from Audience suggestions: controls limit who may see an ad; suggestions guide delivery and can be expanded. Set the right constraints, then fund the creative, conversion signals and holdout tests that can improve
The date matters. Meta announced this class of automation in August 2022, framing it as a way to consolidate campaign setup and let the system find buyers (Meta newsroom, August 2022). Google went the same way with Performance Max, where audience signals are documented as hints that help the system learn faster (Google Ads Help, answer 14530785). The system allocates delivery within the applicable controls. A suggested audience is not a hard restriction, and that distinction should be clear in the
Three inputs the auction obeys, and one that tells you what
Separate required controls from repeated audience tinkering
Placement exclusions and audience controls can be real restrictions, so checking them is useful. The waste is treating every audience suggestion as a precise fence, then repeatedly rearranging it without a testable hypothesis. Google’s audience-signal documentation describes guidance for Performance Max, not a reason to disregard all
Meanwhile creative output holds flat. That is the expensive part, because the auction reprices delivery faster than any planner reads a report, and the only input it cannot infer on its own is a
Every buyer who has swapped a lookalike for broad and watched cost per acquisition hold steady already knows this. The finding is the flat creative line, and it shows up as a miss in the same quarter you
What the platform obeys, and what it treats as a suggestion
Two of the biggest automated campaign types publish their own answer, so start there rather than with a vendor’s
| Input | Meta Advantage+ | Google Performance Max |
|---|---|---|
| Who sees the ad | Delivery within applicable audience controls | Delivery within applicable campaign controls |
| Suggested audience | May expand beyond suggestions; controls differ | Audience signals guide learning rather than restrict delivery |
| Creative | Concepts you supply | Asset groups you supply |
| Bidding | Automated | Automated |
Both columns come from the platforms’ own documentation: Meta on controls and suggestions and Google on audience signals as hints (Google Ads Help, answer 14530785).
Practitioners outside our shop reached the same read in the summer of 2026: the platforms want fewer restrictions, a conversion goal and better creative, which makes the creative the targeting (Kallisto Tech). MediaPost ran the media-side version on June 23 2026, under a headline about advertisers buying screens rather than shows (MediaPost, article 416028).
Document the controls the platform must obey. Revisit suggestions when there is a clear hypothesis and enough data to evaluate it. If you are still budgeting off channel averages, the Google Ads statistics worth arguing with are the delivery ones, not the
One account, with the window attached
Doctrine is cheap, so here is a plan we ran under it. On Sad Panda Studios’ Crush Crush we bought the first cohort, 55,978 installs between 1 and 30 May 2026, on $18,347 of spend across Google and Meta app campaigns, with the audience left to the auction and the money spent on creative supply and a clean conversion signal instead. By day 90 that cohort had returned 252% of its spend in net revenue against a 130% one-year target, and the day-30 forecast we froze at day 7, about 194% blended, came in
The window is part of the claim, not a footnote to it. Adjust attribution, 7-day click and 24-hour impression, blended across platforms, net of store fees, revenue through 10 September 2026, read off the client’s own published dashboard. Move the click window and the same media buy reports a different number, which is why a ROAS figure quoted without one tells you nothing you can
Two honest limits. This is one game in one genre, so it is evidence that the approach can work rather than proof it generalises, and a single cohort cannot separate what the algorithm did from what the creative did. That second question needs a holdout, which is the next section’s argument and the reason we
Fund creative supply, the conversion signal, and the holdout
Three things in the plan are still yours: the concepts you upload, the conversion events you send back, and the calendar of holdout tests. Each needs money attached, and the last two are the lines that get cut first when the
Creative volume is a supply problem. The auction needs enough distinct concepts to find pockets of demand, and distinct means a different promise or a different format. A fourth colour grade of the same hook gives the system nothing to work with. Creative fatigue is the standard failure in automated buying, and the documented fix is refreshing assets on a cadence and testing variations (Measured, media planning guide, 2026).
Auction volume decides how the money splits. A thin account gets little from a holdout, so it should carry a heavier creative line and buy measurement later. A thick account can hold markets out and still hit
The conversion signal is the second controllable and the one most teams underfund. Send back net revenue, after the store takes its cut, or the auction optimizes toward the wrong buyer. Google Play’s fee tiers changed that arithmetic mid-2026, and we broke it down in the Play external payments fee changes. If the purchase events come from a subscription stack, the plumbing choice is a revenue decision; see our read on RevenueCat alternatives.
A holdout only tells the truth above a certain size
A geo holdout needs treatment markets, control markets, and a window long enough for an effect to separate from noise. Without both cells there is nothing incremental to read, which is why measurement has to be a funded line in
Last-click tells you which touch got tagged. Media mix modeling tells you how channels behaved historically at the aggregate level, which is still a live question in the era of automated buying (Measured on MMM). Incrementality tells you what would not have happened without the spend (Measured on incremental sales). Geo-testing on aggregated, non-personal data is what replaced cookie-based reads, and that is why it is now the standard tool (Measured, media planning guide, 2026).
The holdout costs real money in forgone spend. That is the price of the answer, and it is cheaper than a year of defending platform-reported ROAS to a board. We will not run a lift test below the size where it can resolve, and we will not report platform-attributed lift as incrementality. There is no universal spend or percentage threshold that makes a geo test valid. The test needs enough volume, suitable control markets and a duration capable of resolving the effect you care about. If it cannot do that, report the limitation and use cohort payback without calling it proof
The objection: broad targeting re-buys customers you already own
Automated campaigns cannibalize brand demand and re-purchase existing customers at a premium, and the platform counts those conversions as incremental. Scaling into that is how a Meta account posts a great reported return while the blended number goes sideways (Measured on scaling Facebook campaigns).
For an acquisition mandate, review the platform’s supported existing-customer controls and verify how they are configured. Omitting existing-customer purchase events from server-side reporting is not the same as preventing ad delivery to those customers, and it can distort measurement. Keep conversion reporting accurate and distinguish acquisition from retention in
Past a certain spend level the constraint changes again. Frequency climbs, the pocket the auction found gets saturated, and fatigue arrives faster than the production pipeline can answer. At that point the binding limit on growth is how many distinct concepts you can ship per week, which is a staffing and planning problem (Measured on AI in marketing).
What changes on Monday for the media plan
Take the list of proposed changes for the week and sort it into three buckets: does this change creative supply, does this change the conversion signal, or does this change the holdout calendar. Keep required audience restrictions, exclusions, budgets and measurement checks on the review list as well. Automation does not remove responsibility for
For years the audience build was the plan, and the deck opened on segments because that was the part a buyer controlled. Advantage+ and Performance Max took that control back and handed something else over, so the plan now opens on the creative slate and the lift calendar, with the audience section documenting the actual controls, suggestions and customer treatment. Put the geo holdout window on the calendar before the next quarter starts, because a test you schedule after the budget is committed is a test you will not run. The commercial scope for that media plan is paid advertising, including Meta ads when Advantage+ is
Talk